Source Camera Identification Based On Visual Attention And Hierarchical Multi-task Learning

With the development of network technology and digital devices,it has become a habit of people to use cameras and mobile phones to record and share their lives.The production of a large number of pictures on the Internet makes it easier and easier to modify and edit pictures.The widespread use of image editing software enables people to meet their daily needs,but at the same time it also produces some drawbacks.Some lawbreakers use the tampered images in news or laws,which has caused serious damage to the credibility of public organizations.Therefore,detecting tampered images has become an important task in the field of forensics.The source camera identification is proposed under the above background,and its purpose is to determine the specific brand,model,and even equipment of the acquired digital image of unknown source.When the camera takes an image,the process from the real scene to the formation of the image file inside the camera will leave the digital image with the unique noise of the camera.Since the noise patterns of the cameras of the same category are the same,the source camera of the image can be identified as the feature of the source camera.The identity of the image can be verified by recognizing the category of the image source camera,and the determination of the image source camera category can further confirm whether the complete image has been tampered with,and realize the detection of the tampered area.Based on the traditional method of source camera identification,the filtering method is used,and the filtered noise is used as the noise feature of the source camera,and then the template matching is used for classification.The deep learning method is based on the data set training to obtain the model,and automatically learns the characteristics of the source camera.This article uses deep learning to identify the source camera,as follows:This paper proposes a source camera feature extraction method based on multi-scale features and channel domain attention mechanism.This method first extracts shallow features from the data,and uses branches of multiple channels with different convolution kernel scales.The features of multiple channels are merged using self-learning convolutional layers of different scales,and the best scale suitable for this task is determined through experimental comparison.The subsequent network uses residual network combined with the SE module based on the channel domain attention mechanism,adds a weight to each channel of the feature,focuses attention on the important channels,and determines the dimensionality reduction ratio in the SE module through experiments.Compared with the reference residual network,the device-level recognition accuracy rate of the public data set(Dresden data set)is improved by about 1.9%when using this method.Using the hierarchical structure between the categories of the source camera data set,a training strategy of hierarchical multi-task learning is adopted for multiple levels of classification tasks.Taking into account the correlation between the brand,model and even equipment of the camera to form a hierarchical structure,this method first solves the brand classification,and then solves the model classification,until the classification of the most subdivided category.The loss of hierarchical multi-task learning is the weighted sum of the loss of multiple hierarchical classification tasks.The weight ratio between multiple tasks is determined through experiments,that is,when multiple tasks account for the same overall loss weight,the average of multiple tasks is accurate The highest rate.This method can extract the correlation between different levels of categories and improve the recognition effect of each level category.The Dresden data set is used to test the recognition accuracy,which is about 2.3%higher than the baseline residual network equipment level.Limited by factors such as computing resources,instead of inputting the entire picture into the neural network,it is first necessary to extract image blocks from the digital image.This paper uses the image block extraction method based on content texture criteria to extract image blocks from the public camera data set(Dresden data set)and the self-built mobile phone data set,and then trains the image blocks extracted from the complete image in multiple levels.It analyzes its recognition accuracy and confusion matrix,and finally uses the source camera recognition method proposed in this paper to assist in the task of tampering detection.